FEDERAL RESERVE BANK OF PHILADELPHIA

Ten Independence Mall, Philadelphia, PA 19106-1574• (215) 574-6428• www.phil.frb.org

WORKING PAPERSRESEARCH DEPARTMENT

WORKING PAPER NO. 02-16THE RETURNS TO SPEAKING A SECOND LANGUAGE

Albert SaizFederal Reserve Bank of Philadelphia

Harvard University

Elena ZoidoHarvard University

October 2002

WORKING PAPER NO. 02-16

THE RETURNS TO SPEAKING A SECOND LANGUAGE*

Albert Saiz†

Federal Reserve Bank of Philadelphia and Harvard University

Elena Zoido‡

Harvard University

October 2002

* The views expressed in this paper are those of the authors and do not necessarily represent the views ofthe Federal Reserve Bank of Philadelphia or the Federal Reserve System. We want to thank Larry Katz forguidance. David Cutler, Caroline Hoxby, John McHale, Uri Simonsohn, and participants at the HarvardLabor/Public Finance Workshop provided very helpful comments. Bridget Terry Long generously sharedwith us college quality data.† E-mail: Albert.Saiz@phil.frb.org‡ E-mail: ezoido@post.harvard.edu

Abstract

Does speaking a foreign language have an impact on earnings? We use a variety of empiricalstrategies to address this issue for a representative sample of U.S. college graduates. OLSregressions with a complete set of controls to minimize concerns about omitted variable biases,propensity score methods, and panel data techniques all lead to similar conclusions. The hourlyearnings of those who speak a foreign language are more than 2 percent higher than the earningsof those who do not. We obtain higher and more imprecise point estimates using state highschool graduation and college entry and graduation requirements as instrumental variables.

JEL: J31, I29.Keywords: Second language, earnings, human capital.

1

1 Introduction

Does learning a second language in the United States pay off? This is a question that most

Americans face and answer in their education years. It is more relevant as the use of English

as the international common language becomes more extended. Learning a second language is

a possibility many consider at some point. Language courses are an option or a requirement in

elementary, secondary, and college education. Private foreign language institutions, educational

materials, and distance courses are regularly advertised in newspapers and magazines. Reports in

the popular press suggest that speaking a second language has become a “desirable skill in the

work force, one employers are willing to pay for.”1 Despite the salience of the decision of whether

to invest in a second language and the anecdotal evidence that suggests that it is a valuable skill

in the labor market, little research has been undertaken on the labor market returns to learning

a second language in the United States. The literature on language acquisition and labor market

outcomes has so far concentrated on the returns to learning English for immigrants.2 In this paper,

we estimate the returns to speaking a second language for college graduates who are native English

speakers in the U.S.

We find that college graduates who speak a second language earn, on average, wages that are

2 percent higher than those who don’t. We include a complete set of controls for general ability

using information on grades and college admission tests and reduce the concern that selection drives

the results controlling for the academic major chosen by the student. We obtain similar results

with simple regression methods if we use nonparametric methods based on the propensity score

and if we exploit the temporal variation in the knowledge of a second language. The estimates,

thus, are not driven by observable differences in the composition of the pools of bilinguals and

monolinguals, by the linear functional form that we impose in OLS regressions, or by constant

unobserved heterogeneity. To reduce the concern that omitted variables bias our estimates, we make

use of several instrumental variables (IVs). Using high school and college graduation requirements

as instruments, we estimate more substantial returns to learning a second language, on the order

of 14 to 30 percent. These results have high standard errors, but they suggest that OLS estimates

may actually be biased downward.

The estimation of the returns to speaking foreign languages may have important policy impli-

1 New York Times, October 18, 1999.2 Some examples are Carliner (1995) and Chiswick and Miller (1997b, 1998). Section 3 discusses this literature in

more detail.

2

cations. Many states regularly update their high school graduation requirements. There is much

discussion on the skills that secondary education should provide. Skills rewarded in the labor market

seem of paramount importance among these. The relative labor market value of second language

skills should thus weigh heavily in these choices. Colleges and other educational institutions also

need to make decisions on graduation requirements. This paper should provide helpful input to

these issues.

Furthermore, our results may help us understand individual decisions on whether to learn a

second language. Human capital theory guides our understanding of the individual’s decision. One

should invest in the acquisition of a foreign language if the present value of the future returns for

doing so exceeds the costs. Some of the returns from learning a second language consist of the

direct consumption of services produced by the individual’s knowledge of the language. Speaking

a second language while travelling abroad, asking an immigrant shopkeeper for a product in her

native language, and relating to foreign friends are all examples of these. This paper cannot address

them. The value of the labor market skills that learning a foreign language provides is the other

important component of the returns to learning the language. Our objective is to quantify it.

There are several reasons that can make the knowledge of a foreign language useful in the

American labor market. First, American companies export and import products to the rest of the

world. Knowledge of a second language may be a valuable asset for such companies doing business

abroad or for companies catering to immigrants or people of foreign background within the U.S.

Offering a service in the language of the prospective customers may provide a competitive advantage

for firms, especially in areas where the foreign born tend to settle. Finally, second language skills

are in demand from several government agencies that deal with foreign affairs: diplomatic service,

CIA, military intelligence, and the like.

Additionally, an extensive literature suggests that learning a second language may help indi-

viduals develop their cognitive and communicative abilities. Research in linguistics underlines the

possible advantages of bilingualism in terms of intellectual and academic achievement.3 If speak-

ing a second language is important for improving cognitive capabilities, we should find that the

individuals who speak a second language are more productive and earn higher wages. At the same

3 Cooper (1987) finds that math and verbal scores are higher with each additional year of foreign language study.Olsen and Brown (1992) show that “students who had completed a foreign language course in high school tendedto have higher scores on the ACT exams in English and math regardless of their ability level.” Learning a foreignlanguage may help develop analytic and interpretative capacities. Cook (1997) reports that “increased metalinguisticawareness of phonology, syntax, and the arbitrary nature of meaning, and gains in cognitive flexibility” are establishedoutcomes from learning a second language. This literature, though, faces challenges similar to the ones we face inour research. Section 3 presents a detailed discussion of these issues.

3

time, access to foreign media and literature may help innovation and adoption of best practices

from abroad and improve workers’ productivity.

The paper proceeds as follows. In Section 2 we discuss the existing literature on the returns

to learning a second language. Section 3 addresses the potential methodological problems that

estimating the returns to a second language faces. In Section 4, we introduce the data sources

and describe the characteristics of individuals speaking a second language in our sample. Section

5 presents the results of the OLS, propensity scores, panel data, and IV approaches. Section 6

concludes.

2 The Literature on Second Language and Earnings

Economic analysis has been applied only recently to our understanding of the individual decision

to invest in a second language. Grenier and Vaillancourt (1983) were the first to identify foreign

languages as an element in the human capital portfolio of individuals. They use the framework

laid out by Becker (1964) and Mincer (1974) to describe how individual characteristics have an

impact on the gross costs, the gross returns and the information on the value of the returns to

this investment. Similarly, Ridler and Pons-Ridler (1984) analyze the decision to learn a second

language as an investment decision. These authors emphasize the importance of second languages

as consumption goods and discuss the economic costs of forced “francisation” in Quebec as an

economic problem.

So far, the empirical literature on the returns to speaking a second language has concentrated

on the incentives and wage premia that immigrants receive for mastering the language of their

country of adoption. Carliner (1995) finds that most immigrants in the U.S. are proficient in

English. He shows that each additional year of residence in the U.S. increases the probability of

being proficient in English by 1.1 percentage points. Chiswick and Miller (1997a, 1998) show that

the probability of speaking English for immigrants in the U.S. responds to the costs and benefits of

doing so. Chiswick and Miller (1997b) study the labor market returns to English acquisition among

immigrants in the United States. They find that foreign-born non-English speaking immigrants

who are fluent in English earn about 15 to 19 percent more than the ones who are not. Similar

wage premia are observed in other countries. Chiswick (1997), for example, analyzes the wage

differential for immigrants who speak Hebrew in Israel. In Canada, Chiswick and Miller (2000)

find that immigrants who can speak one of the official languages earn wages that are 10 to 12

4

percent higher. They find evidence of positive selection: individuals with higher tendency to learn

an official language tend to earn higher wages, possibly for reasons other than such proficiency.

Despite the extensive literature on the cognitive and developmental effects of bilingualism, there

is little research on the returns to speaking a second language for nonimmigrants.4�5 An important

and motivating finding for us is the fact that foreign language courses in high school do affect wages

(Altonji, 1995). Indeed, his results show that foreign language courses have higher returns than

courses in mathematics, science, and verbal skills. Such strong effects of taking second language

courses on future wages seem puzzling in the American context. These results could be driven by

the correlation between taking second language courses and omitted variables, such as high school

quality or family background, or by selection. One would expect, for example, that foreign language

courses are more sensitive to the affluence of the families whose children attend a school than math,

science, or language courses are. A second possibility, one that we are interested in exploring, is

that speaking a second language is a valuable skill in the labor market.

Finally, and closest to the spirit of our work, Lopez (1999) estimates the labor market returns

to speaking a second language in the U.S. He uses observations form the National Adult Literacy

survey (NALS) of 1992. The main shortcoming of this data set is that only individuals belonging to

a language minority were questioned regarding their language skills. Native English speakers who

have subsequently learned a foreign language, the treatment group of interest for us, are excluded

from the questionnaire on second languages. Thus, this author’s basic test is whether language

retention among language minorities is associated with higher or lower labor market earnings. He

finds that those who speak a second language proficiently earn wages that are 14.2 percent higher

than those earned by people in language minorities who do not speak the minority language well.

He also finds that minority language individuals who speak their mother tongue well earn 13.5

percent more than English monolinguals, but the difference is not significant.

4 See Cook (1997) for a survey on the benefits of bilingualism on the development of cognitive ability and com-munication skills.

5 In the Canadian context, Shapiro and Stelcner (1997) and Christofides and Swidinsky (1998) explore the returnsto speaking a second official language. Pendakur and Pendakur (1998) estimate the returns of speaking French in Van-couver and Toronto and of speaking English in Francophone Montreal. These authors find that speaking a nonofficiallanguage is associated with a negative earnings differential. The interpretation of these results is subject to severalproblems. Selection into learning a language different from the official ones may affect these estimates dramatically.For example, if immigrants speak other languages in much greater proportions than the general population, it will bedifficult to disentangle the effects of assimilation from the pure effects of speaking more languages. This paper pointsto the fact that selection into learning a second language can be correlated with individual attributes that decreaseearnings, which suggests the importance of controlling for potentially confounding factors.

5

3 Empirical Strategy

Measuring the returns to learning a foreign language presents considerable methodological chal-

lenges, in common with the work on the returns to schooling. Think of a running a regression like

the following:

log�� = �+ ��� + ���� + ��

where log� is the log of earnings, � is a vector of personal characteristics, and �� is an indicator

of whether the individual speaks a second language.

Obviously, speaking a second language is not randomly assigned in the population. Individuals

choose whether to learn a foreign language according to their potential earnings. Thus, a selection

problem arises. Consider the distribution of the benefits of speaking a second language among the

population. Assume that the costs of learning it are the same for all individuals. Then, those

individuals with higher returns to speaking a second language will have greater incentives to invest

in learning it. Empirically, we may find that the individuals who speak a second language obtain

higher earnings. The average returns of speaking it, if the second language skills were randomly

assigned, would be smaller. Selection by earnings will, thus, bias conventional estimates upward.

This problem is akin to the selection problem in the labor market participation decision of female

workers (Heckman, 1980).

Additionally, the costs of learning a second language may be smaller for individuals with higher

cognitive competence. If this is the case, more able individuals will learn foreign languages with

a higher probability. If we do not correct for this differential selection with respect to ability, our

estimates will contain an ability bias, as the coefficient on the foreign language variable will capture

part of the effect of ability on earnings (Griliches, 1977). Again, this causes the estimates of the

returns to a second language to be biased upward.

We also need to take into account that there may be other unobservable characteristics of the

individuals that simultaneously determine earnings and whether an individual speaks a second

language. People who study foreign languages may have different preferences regarding the types

of jobs they like and the kind of studies they choose. Consider the population of college graduates.

People in humanities majors study foreign languages with higher likelihood and are more likely to

choose teaching jobs, for example. This selection problem would again bias our estimates, although

the direction of the bias is not clear a priori.

An additional problem of the OLS specification is that it imposes a linear relation among

6

the variables. The returns to learning a second language, however, may vary markedly among

different groups. If the effect of the treatment we are trying to estimate (the effects for those

who learn a foreign language of doing so) is contingent on observable individual characteristics,

simple OLS regressions may not yield unbiased estimates. We can think of this as a problem

of model specification. The coefficient of the foreign language indicator cannot be interpreted

independently of the other characteristics of the individual. A fully interacted model may not be

the solution because we do not know ex-ante the relevant interactions and the functional form of the

treatment effect. Moreover, if the treatment group (the people who speak a second language) and

the control group (those who do not) are different in their observable characteristics, the coefficient

of a treatment indicator may be very sensitive to the specification (Lalonde, 1986).

We address all these concerns through several complementary approaches in our empirical im-

plementation. First, we include a very complete set of control variables in our OLS regressions to

reduce the concern that selection may bias our estimates. We try to mitigate the ability biases by

including SAT scores, GPA, parental education, and indicators of the quality of the college attended

in our regressions. We include controls for the major chosen by the individuals to capture unob-

served individuals’ characteristics, as the choice of a college major may be viewed as an indicator

of career preferences. Second, we address the possibility that the linearity imposed in the OLS

regression biases our estimates of the average effect of speaking a second language on earnings. We

examine the sensitivity of the results to this assumption by using several methods based on the

propensity score: weighting, stratification, and matching.

We will be able to give a causal interpretation to the estimates obtained by OLS regression

and propensity score methods only if, once we condition on the observable characteristics of the

individuals in our sample, speaking a foreign language or not is independent of an individual’s

potential earnings. This may not be the case. To address the problem of constant unobserved

individual characteristics that may bias our results, we exploit the longitudinal dimension of the

data set. We thus compare the evolution of earnings for people who learned a second language

between 1993 and 1997 to the evolution of earnings for other groups. Finally, we address the

“selection by earnings” problem (and the rest of omitted variables that we may have not succeeded

in controlling for) using instrumental variables (IV). High school foreign language requirements in

the state where the respondent attended high school and college second language requirements are

the instruments used.

7

4 Data

We focus on the returns to speaking a second language in the U.S. There are no available data

sources that contain information on foreign language proficiency for a representative sample of

the American population. Our main data source is the “Baccalaureate and Beyond Longitudinal

Study” (B&B henceforth), from the National Center for Education Statistics, which contains unique

information on second language ability. This data set tracks the experiences of a nationally rep-

resentative cohort of college graduates who received their bachelor’s degrees during the 1992-1993

academic year. It contains information on the characteristics of the universities the individuals

attended (four-year institutions in all cases), from the students’ undergraduate transcripts, and

from answers to several questions posed to the individuals in surveys carried out in 1993, 1994, and

1997. These questions include information on demographics, parents’ background, earnings and

job characteristics, academic major, and courses taken in college and after graduation. The sample

included 11,192 students initially. Of these, 9,274 individuals responded to the initial interview and

all follow ups.

Respondents were asked in 1993 and 1997 whether they spoke a foreign language and to identify

which language. The question used was, “Do you have conversational knowledge of languages other

than English?” If the answer was affirmative, the interviewer followed up with the question “What

are these languages?” This question is ideal for us, as it is intended to capture knowledge of a second

language and not whether an individual belongs to a language minority. This feature contrasts with

other data sets, which contain questions on second language only for linguistic minorities.

Our choice of this data set is therefore very much driven by the existence of a second language

question geared toward native speakers. Other features of B&B, however, make it very suitable

for this paper because it includes variables that can be considered cognitive ability proxies, namely

SAT scores and college grade point average (GPA). It provides information on where the parents

of the individual were born, whether the respondent spoke English at home while growing up, and

whether she is an American citizen. We will use these variables to focus our analysis on individuals

born to parents born in the U.S., American citizens who spoke English as their mother tongue.

B&B also contains a wealth of information on academic and job careers and on earnings. The only

shortcoming of our using this data source is that the results we obtain are representative only for

the population of college graduates. Thus, we suggest caution with their generalization. It is not

clear, a priori, if one should expect greater or smaller effects on the general population.

8

We complement B&B with some additional information. We use an indicator of college quality

compiled by Caroline Hoxby and Bridget Terry Long.6 This is a classification of colleges into six

categories according to the average SAT of admitted students in the college and the average high

school GPA of attending students. Per capita income in the states of residence in 1989, 1993,

and 1997 is from the Bureau of Labor Statistics. The percentage of Hispanic residents in a State

are obtained for 1989, 1993, and 1997 from the U.S. Census Population Estimates, available at

www.census.gov.

In our IV estimation we exploit variation in high school and college graduation requirements.

High school graduation requirements in 1989 (the estimated year for which our college seniors

graduated from high school) are obtained from the Digest of Education Statistics. We gathered

information on second language college requirements from the Modern Language Association (MLA)

Survey of Foreign Language Entrance and Degree Requirements (FLEDR) (Brod and Lapointe,

1989). The FLEDR was undertaken in July 1988. Questionnaires were mailed to institutions of

higher education that offered courses in one foreign language or more. The response rate was 98.3

percent. The survey contains data on whether an institution has foreign language admission and

graduation requirements and which majors and degrees (e.g., BA or BS) require foreign languages

for graduation. The survey also includes the number of semesters or semester hours of a second

language required for graduation, if any. We concentrate on graduation requirements and focus on

whether an institution has any foreign language requirement. This avoids endogeneity problems,

as opposed to using an indicator on whether the major and degree of the student in the sample

required a foreign language: people self-select into majors and degrees. We use these variables

as instruments that should help predict whether a college graduate has knowledge of a second

language.

The basic demographic characteristics of the individuals in our sample are presented in Table

A.1. in the Appendix. The table is based on the 1997 wave of the survey. Panel A shows that the

average individual is around 30 years old in this second survey and that she has more than 3 and a

half years of labor market experience. A third of the individuals are married; less than half of them

are men. More than a quarter of the individuals have received some graduate degree, and most

of them (67 percent) attended a public college to obtain their degree. Thirty-four percent of the

individuals in the sample claim to speak a foreign language or more.7 Panel B shows the regional

6 See Appendix A and Hoxby and Long (1999).7 The proportion is slightly lower (28 percent) in the subsample of individuals who spoke English at home while

growing up, whose parents were born in the US and who are American citizens.

9

distribution of the individuals in the sample.

4.1 Who Speaks a Foreign Language?

We start our analysis with a brief description of the characteristics of the individuals who speak

one or more foreign languages. Table 1 compares individuals who speak and who do not speak

foreign languages. Panel A reveals some differences between these two groups: the probability of

working, having attended a public college, holding an MBA degree, and being married is higher

among those who speak English only. They are older and have slightly more work experience.8 On

the other hand, those who speak a foreign language are more likely to hold other graduate degrees,

are disproportionately more likely to be black, and have slightly higher wages.9 Interestingly, those

who speak a language other than English are relatively more likely to have high SAT scores (Panel

B). This suggests that those who decide to learn a foreign language may have higher cognitive skills.

The proportion of residents who speak a foreign language varies by region, as Panel C in Table

1 shows. New England and the Pacific region have the highest proportion of residents who speak a

second language (41 percent). The percentage of foreign language speakers is lowest in East North

Central (IA, KS, MN, MO, NE, ND, SD) and East South Central (AL, KY, MS, TN).

Note that more than 15 percent of those who speak a foreign language actually speak more than

one language other than English (Table 2, Panel A). Around 14 percent speak two foreign languages,

almost 2 percent speak three. We find no differences by gender in the number of languages spoken.

In Panel B, we observe that most of them speak Spanish (58 percent), followed by French and

German. Four percent of the individuals in our sample speak an eastern or southeastern Asian

language, and 2 percent speak Chinese.10 Spanish remains the most popular language in all regions

and for both men and women. In relative terms, a higher percentage of women choose French

and Italian as their second language, while men prefer German (the differences are statistically

significant). Different languages are spoken by residents of different regions: Spanish is relatively

more popular as a foreign language in the West South Central, East South Central, and Mountain

8 Experience is calculated by, first, obtaining the months between graduation and the interview date. We thensubtract the months of full-time graduate study that the interviewee reports.

9 We obtain hourly wages, the dependent variable of interest to us, by dividing yearly, monthly, weekly, or dailywages by the total number of hours the individual reports to work in the relevant period. We find some extremevalues in earnings and adjust the sample by truncating the sample at the 1 percent. That is, we drop from the samplethe observations in either tail of the distribution. See Angrist and Krueger (1999) for an overview of these trimmingmethods as a way to reduce the biases due to measurement error.10 The proportions are very similar in the subsample of native English speakers with parents born in the U.S.

63 percent speak Spanish, 24 percent choose French, 12 percent German. Only 0.6 percent speak an Eastern orSoutheastern Asian language, and 0.4 percent speak Chinese.

10

regions; residents in the South Atlantic region and New England speak French more frequently,

while German speakers are disproportionately located in the West North Central region.

5 The Returns to Speaking a Second Language

5.1 OLS Results

We begin our investigation of the returns to speaking a second language with a variety of conven-

tional earnings functions estimated by OLS. Table 3 reports the estimates. We include quadratic

functions of age and experience, and race, gender, marital status and parental education controls.

We add controls for the income per capita in the state of residence11, for college quality, for whether

the individual holds a graduate degree (MBA, Ph.D., and other master’s) and ability proxies (nor-

malized GPA and SAT quartile)12. It is particularly important to account for the major in college.

In appendix Table A.2 we illustrate how graduates from majors in which the percentage of indi-

viduals who report speaking a second language is higher are also the majors in which the average

wages are lower.

The estimated foreign language coefficient implies a 2.8 percent wage premium associated with

speaking a foreign language for the average individual in the sample (column 1). We estimate

additional models to test the robustness of this estimate. In particular, we restrict our sample to

those individuals whose native language is English and account for the possibility that some of

the individuals in our sample might be part-time workers. In the regressions shown in the next

two columns, we restrict the sample to individuals who spoke English at home while growing up

(column 2). Column 3 focuses on individuals who spoke English at home while growing up, are

American citizens, and whose parents were born in the U.S. (more than 80 percent of the sample).

Finally, column 4 shows the returns for full-time workers who spoke English at home when they

were children.13 Again, the point estimates are similar, although slightly lower: 1.9-2.2 percent.

These compare with estimates in the range of 0.9 percent when we exclude the controls for the

college major chosen by the individual.

In separate (unreported) regressions, we explore the labor market returns to speaking specific

languages. We estimate OLS regressions following the previous specifications but allow the coeffi-

cient to vary by language spoken. In our sample, German is the language that obtains the highest

11This yields the same results as the addition of state of residence dummies.12Results do not change if we control for industry and occupation dummies.13 We define an individual as a full-time worker if she works more than 35 hours a week.

11

rewards in the labor market. The returns to speaking German are 3.8 percent, while they are 2.3

for speaking French and 1.5 for speaking Spanish. In fact, only the returns to speaking German re-

main statistically significant in this regression. The results indicate that those who speak languages

known by a smaller number of people obtain higher rewards in the labor market.14

We then check whether the returns to speaking foreign languages are homogeneous across the

population. In particular, we analyze whether the returns differ for groups with different charac-

teristics. We examine differences by gender, race, grades in college and graduate degree, by the

percentage of Hispanics in the state of residence (for those whose second language is Spanish), and

by occupation. The results (not reported in this version of the paper) reveal that speaking Spanish

pays off less in states in which larger shares of the population are of Hispanic origin (probably

because of a supply effect). By occupation, individuals in business support, management positions,

or occupations related with science and technology are the ones who are more highly rewarded

in the labor market for their foreign language skills. We do not find any statistically significant

differences for the other individual characteristics analyzed.

These results should be interpreted with caution. In our regressions, we include a complete set of

controls for ability (GPA and SAT scores), marital status, college quality, and parental background,

to try to reduce the concern of omitted variable biases. Still, we acknowledge the possibility that

unobservables correlated with wages and the ability to speak foreign languages may bias our OLS

estimates. We therefore use information on changes in the ability to speak a second language

to examine whether the estimates from our baseline earnings equations simply capture an ability

bias. Our strategy is the following. If learning a second language requires cognitive skills that are

not captured in our baseline specifications, the coefficient on the second language indicator may

be capturing the effect on earnings of some of these. We can exploit the differences between the

responses regarding second language knowledge between the 1993 and 1997 interviews to examine

whether people with the necessary ability to learn a second language, but who could not speak it

in 1997, earn the same wage premium as second language speakers.15

14 This is consistent with anecdotal evidence. See, for example, Wall Street Journal, November 13, 1999: “Wherebilingual workers are in short supply, the employees may command a premium or land a job they wouldn’t otherwisehave gotten.”15 The transition matrix in and out of speaking a second language between 1993 and 1997 is the following:

Speaks FL in 1993 Does not speak FL in 1993 Obs:Speaks FL in 1997 2621 677 3298

Does not speak FL in 1997 1103 4947 6050Observations: 3724 5624 9348

12

In Table 4 we replicate the previous OLS regressions, but now we add a dummy that equals

one if the individual can speak a second language in 1993 but not in 1997. If the cognitive ability

necessary to learn a second language is the only determinant of earnings, we should see a similar

coefficient for those who could speak a second language in 1997 and those who were able to speak

it in 1993. In fact, the coefficient on the indicator for speaking a second language in 1993 only is

not very different from zero and is statistically insignificant. We proceed with the same estimation

for the subsamples of native English speakers and of native speakers who are American citizens

and whose parents were born in the U.S. We do so in order to separate the dynamics of those who

learn a second language from those of language minorities because the different composition of the

two groups could, in principle, make the interpretation of the results difficult. The results confirm

that the returns to a second language are not completely driven by an ability bias.

These estimates exploit the fact that some individuals decide to learn a language and do not

subsequently exert an effort in keeping it alive. The returns of speaking the language for the latter

may have been very low. An unbiased estimate of the ex-ante returns to speaking a second language,

taking into account that some forget it, is given by the difference between the coefficient of those

who speak minus the coefficient of those who forgot, multiplied by the fraction of people who speak

now out of those who ever spoke a second language.16 Such estimates yield an approximate return

of 2 percent.

5.2 Propensity Score Methods

If the returns to speaking a second language vary depending on individual characteristics, or if the

people who speak and do not speak a second language are very different, as we saw in Table 1,

our estimates of the returns to speaking a second language may be biased because of the linearity

imposed in OLS regressions. To explore how sensitive our results are to the linearity assumption, we

opt for propensity score methods to give flexibility to the functional form of the effect on earnings

of speaking a second language across groups.

The propensity score is the probability of being assigned to a treatment, conditional on a set of

covariates. The treatment, in our setting, is speaking a foreign language. The estimated propensity

score allows us to control for differences between treatment and control groups when the treatment

is not randomly assigned, the number of predetermined variables is high, and the groups are not

very similar. It conveniently summarizes all individual characteristics in a single variable in the

16 See Appendix B for an extensive discussion.

13

unit interval. Under selection on observables, conditioning on the propensity score is enough to

have independence between the treatment indicator and the potential outcomes (Rosenbaum and

Rubin, 1983).

While the assumption of selection on observables may seem strong in this setting, Dehejia

and Wahba (1999) show that propensity score methods yield estimates of treatment effects that

are closer to experimental benchmark estimates than traditional econometric techniques for non-

experimental data.17 Thus, although the propensity score is unknown and has to be estimated,

inference for average treatment effects seems to be less sensitive to specifications of the propensity

score than to the specification of the conditional expectation of potential outcomes implicit in OLS

regressions.

We first estimate the propensity score for our sample using a logistic probability model. Table

A.3 in the Appendix presents the coefficients from this estimation.18 Conditional on the propensity

score, the covariates should be independent of the assignment to the treatment, i.e., of whether

the individuals speak a foreign language. We grouped the observations into strata defined on the

propensity score and checked whether the covariates were balanced within each stratum. We did not

find significant differences in the distribution of the covariates within these groups. After estimating

the propensity score, we use matching, weighting on the propensity score, and stratification to

produce alternative estimates of the returns to speaking a second language.

With stratification, we begin by sorting the observations according to their estimated propensity

score. Panel A in Table 5 shows the differences in average earnings between treatment and control

units within the propensity score deciles. We estimate the treatment effect by adding the within-

strata differences in earnings and weighting by the number of treated observations in each decile.

The estimated effect of speaking a second language is 1.2 percent for those who speak (the selected

average treatment effect, or SATE). We also obtain an alternative estimate after adjusting for

covariates, to eliminate any possible remaining within-block differences. One would expect this

17 These authors compare a randomized evaluation of the National Supported Work (NSW) Demonstration withnon-experimental comparison units from survey data sets, following Lalonde’s analysis of this issue in 1986. Theyconclude that “there may be important unobservable covariates for which the propensity score method cannot account.However, [...], there are substantial rewards in exploring first the information contained in the variables that areobserved. Propensity score methods can offer both a diagnostic on the quality of the comparison and a means toestimate the treatment impact.”18 Note that we are conditioning on pre-treatment variables only (gender, race and family background). Rosen-

baum (1984) shows that if the treatment is only ignorable given some unobserved pretreatment variables, propensityscore methods may yield biased estimates if they condition on observed pretreatment variables. If this is the case,adjustment for pretreatment and some posttreatment characteristics may yield unbiased estimates. We thereforeexplored alternative specifications of the logistic model that included post-treatment individual characteristics. Inresults not reported here, we found that the estimates of the effect of the treatment on earnings remained unchanged.

14

adjustment to have little effect if the covariates are well balanced within the strata. Indeed, our

estimate of the effect of speaking a second language on wages increases only slightly, up to 2 percent.

In the second method based on the propensity score, we match each individual who speaks

a foreign language with the comparison individual with the closest propensity score.19 We then

drop from our sample the observations that correspond to individuals who do not speak a second

language and were not matched to “treated” individuals. Table A.3 in the Appendix presents the

characteristics of the matched sample for those individuals who speak a foreign language and those

who do not. For this matched sample, we estimate a return (SATE) to speaking a second language

of 2.8 percent (Panel B, Table 5). We obtain an adjusted estimate of the returns to speaking a

second language including covariates in the regression, to account for any remaining observable

differences between treated and control individuals. The point estimate increases slightly, up to 4

percent.

A final method of exploiting the propensity score is through weighting by the inverse of the

probability of receiving the treatment (Horvitz and Thompson, 1952). Using this technique, the

point estimate remains close to those obtained with the previous methods, 1.1 percent. Standard

errors increase relative to other methods (.098) and make it impossible to reject a zero return to

speaking a foreign language.20 This finding is not surprising, though, as this is a relatively inefficient

estimating method.21

All propensity score methods confirm our findings from the OLS regressions: speaking a foreign

language is a valuable skill in the job market. Its returns are around 2-4 percent. While the

identification conditions that OLS and propensity score methods require for the estimates to be

unbiased are the same, the approach that we explore in this section helps us construct estimates of

the conditional expectation of potential outcomes without the functional form restrictions that OLS

regression imposes. The fact that all the estimates point in the same direction and are consistent

across specifications and methods is reassuring and confirms our previous conclusion that speaking

a second language is associated with a wage premium in the labor market.

19 We perform matching with replacement, that is, allowing the same observation for a control individual to providethe match for several treatment units. See Dehejia and Wahba (1999).20 As the distribution of standard errors is not analytically tractable, we obtain standard errors using bootstrapping.

Parameters were estimated 200 times to obtain an estimate of their variance. Each of the samples used for theestimation is constructed sampling the original sample with repetition.21 See Hirano, Imbens, and Ridder (2000).

15

5.3 Panel Data Methods

In this section we use the information on earnings and self-reported foreign language skills from

two consecutive surveys to examine if the observed wage profiles of those who learn and forget a

second language between the two surveys are consistent with the results obtained with OLS and

propensity score methods. B&B contains information on second language ability in 1993 and 1997

and on wages in its 1994 and 1997 waves. The 1994 interview takes place when college graduates

have just started their first job after graduation.

We exploit the longitudinal aspect of the data set to address the omitted ability problem inherent

in the previous estimation methods. Consider the following explanatory model for the logarithm

of wages:

ln��� = �� + ����� + �0��� + + ���

where � and are subscripts for individuals and time, respectively. ���� is a dummy that equals 1

if individual � speaks a second language at time and ��� is a vector of explanatory variables. � is

the coefficient of interest. can be interpreted as a time effect, common to all individuals. �� is an

individual fixed effect. One can broadly interpret this coefficient as individual unobserved constant

heterogeneity like ability. The problem with our OLS specification is that we cannot observe ��.

If ���� and �� are correlated, the coefficient on the second language indicator may be partially

capturing the effect of ability on earnings. This suggests using the differences between the values

of the variables in the two periods for which we have information:

(ln��1 − ln��0) = �(���1 − ���0) + �0(��1 −��0) + + (��1 − ��0)

Table 6 shows the results of the estimation of this model. Note that only those variables that

experience any change between both surveys appear in differenced form. The results that we obtain,

in column one, are very similar to our OLS estimates. The wages of people who learn (forget) a

second language tend to increase by almost 2 percent more (less) than other individuals.

One may worry that people who have the ability to learn a second language differ in their wage

level and also in their wage profile over time. If this were the case, our estimate of � in this first

specification would be capturing the steeper wage profile over time for higher ability individuals.

One would then expect that those who speak a second language in both periods (call them always

speakers, AS) should also experience higher wage increases. In column 2 of Table 6 we include

an AS indicator to account for this possibility. As the results show, people who spoke a second

language in both 1993 and 1997 do not experience higher wage increases over time. In column 3 we

16

repeat this exercise for the subsample of individuals who spoke English at home while they were

growing up, are American citizens, and have American-born parents. Again, we want to render the

groups of those who learn a second language later in life and the AS as similar as possible. Our

estimate of � varies little across these specifications. These results thus suggest that the intrinsic

knowledge of a second language is the explanation for the wage hike captured in the panel estimate

of �� As in our OLS regressions, the results are consistent with a labor market return of around 2

percent of learning a second language.

Another possibility is that the wage profile (the steepness of wages with respect to seniority)

may differ according to the observable characteristics of the individuals:

ln��� = �� + ����� + �0��� + + 00��0(1− ) + 01��0+ ���

If the characteristics of the individuals who learn and forget a second language between the two

surveys are different, our estimate of � would capture only such differences. Taking differences:

∆ ln��� = �∆���� + �0∆��� + + 0��0 + ���

where is 1 − 0. In columns 4 and 5 of Table 6 we estimate this model. We control for the

1994 values of the individual characteristics that we used in our baseline specification as regressors.

Column 4 includes all individuals in the sample and column 5 is limited to those English-speaking

natives, third-generation Americans. The results are in line with those in previous specifications.

There are two caveats in the interpretation of the previous estimates. The first is the common

problem of false transitions. Many people who spoke a second language in 1993 and declare they

couldn’t speak it in 1997 may actually be capable of doing so after some short training. Some of

them may not have been capable of speaking it in the first period. People who declare they have

learned a second language may have actually spoken it in the first wave but were not sure about

their answer. It is likely that the measurement error in the question “Do you have conversational

knowledge of languages other than English?” is larger for the group who do not give a consistent

answer in 1993 and 1997 than for the group that did not change its response. This amounts to

bigger measurement error in the variable capturing whether the individual’s knowledge of a foreign

language changes between 1993 and 1997. This should bias our estimates of � downward.

A second problem arises from the selection of individuals into learning and forgetting a second

language. If our results are not driven by ability bias, learning a second language is, indeed, a

profitable investment. Thus, we should expect people with the lower returns to speaking a second

17

language not to make the effort toward its maintenance. Conversely, those learning a second

language after college are most likely the people with higher labor market returns to speaking it.

These facts make the interpretation of our results in Table 6 difficult. More specifically, the returns

for those who learn a second language should be higher than the returns for those who forget it.

Our specification in Table 6 imposes symmetry in the returns for both groups. In Table 7 we

replicate the structure of the previous regressions, but we now allow for different returns for the

two groups. Note that the returns for those who learn and those who forget a second language

can be interpreted as a higher and lower bound for the returns to learning a second language. As

we expected, people who learn a second language between the two B&B surveys have a higher

estimated return (almost 4 percent) than people who report having forgotten it (0 percent). The

results confirm that previous estimates capture an actual labor market value of speaking a second

language, and give us an interval between 0 percent to 4 percent for its return.

5.4 Instrumental Variable Methods

We finally exploit exogenous variation in the decision to learn a foreign language. A consistent

estimate of the returns to speaking a foreign language can be obtained if there is a variable that

affects whether an individual speaks a second language but which does not directly affect earnings.

Note that the interpretation of the IV estimates, however, will differ slightly from the coefficients

we have previously estimated. If the instruments are valid, and if the returns to speaking a foreign

language vary across the population, we will obtain a consistent estimate of the Local Average

Treatment Effect (LATE).22 LATE is the average return for the population of compliers, that is,

those individuals who learn a second language because they are required to do so, and who would

not have learned it if not required. In principle, the economic return might be different for this

group of people than for the average individual in the population (the average treatment effect,

ATE) or the average individual who speaks a foreign language (SATE).

We use high school and college graduation requirements as determinants of whether an indi-

vidual speaks a foreign language that can be excluded from the earnings equation. In this section

we first discuss the validity of the instruments and then present our IV estimates of the returns to

speaking a second language.

22 See Angrist and Imbens (1994).

18

5.4.1 Are Graduation Requirements Legitimate Instruments?

Most of the individuals in our sample graduated from high school in 1989. At the time, several states

required high school graduates to complete a certain number of courses in foreign languages (District

of Columbia, Rhode Island, and Texas) or included foreign languages among several elective subjects

students could choose among in order to complete the necessary number of credits to obtain their

high school diploma (California, New Hampshire, Oklahoma, Oregon, Virginia and West Virginia).

Similarly, some colleges require all or some of their students to satisfy foreign language requirements,

either by taking a certain number of courses or by demonstrating proficiency in a foreign language

test. Table A.5 in the Appendix lists the states and their graduation requirements.

Unfortunately, B&B does not report information on the states where the individuals in the

sample completed high school. We have information, however, on the state where their parents

lived in 1993. We therefore assume that the parents have not moved to other states since their

children graduated from high school, so that the individuals we observe completed high school in

the states where their parents lived in 1993.23 In Panel A of Appendix Table A.6 we compare

individuals whose parents live in states that had foreign language requirements to graduate from

high school graduation with those that do not . We observe some differences between these two

groups’ hourly earnings, the likelihood of speaking foreign languages, of attending a public college

(higher in states with requirements), and of being married (lower in states with requirements).

States with foreign language requirements had slightly higher income per capita in 1989, when the

average individual in our sample was graduating from high school.

We perform the same comparison between individuals who were subject to foreign language

requirements in college and individuals who were not (Table A.6, Panel B.1). Some differences

arise between these two groups in age, experience, marital status, and race. Moreover, colleges

with requirements are located in states with lower per capita income. The proportion of colleges

with requirements varies with quality, and the lowest proportion of colleges with requirements is

found among the least competitive ones. The relation between quality and requirements is non-

monotonic, however: there are more colleges with requirements in the “very competitive” category

than among the “most competitive” ones, for example. Public colleges emphasize foreign language

requirements more than private ones.

23 It is important to note that high school graduation requirements in 1989 are likely to be relevant for theindividuals who actually completed high school in that year. As the summary statistics for our sample revealed, thereis some dispersion in the age distribution. The results in this section remain unchanged, though, if we focus on asubsample of individuals born between 1968 and 1972.

19

These graduation requirements will be legitimate instruments if they affect an individual’s de-

cision to learn a foreign language but have no direct effect on earnings. The observed differences

between individuals affected or not by the requirements suggest that we should include these vari-

ables as controls in our regressions. To control for the possibility that families that place a strong

emphasis on education choose to live in states with stronger high school graduation requirements,

we include the set of requirements in other subjects as controls in our regressions.24 We think it

is safe to assume that foreign language requirements per se do not drive the residential decisions

of families in search of better school systems. The same reasoning drives us to include a complete

set of college quality indicators, to reduce the concern that higher quality colleges impose stronger

foreign language requirements on their students. Once we include these controls, we believe that

graduation requirements are legitimate instruments and that they are not correlated with unob-

servable determinants of individuals’ earnings. We find it hard to argue that these requirements

drive the individual’s or her family’s decision to attend a specific college or to live in a certain state.

5.4.2 Instrumental Variables Estimates

We use high school and college graduation requirements and their interactions as our instruments.

We show the results from the first stage estimation in Table 8. Both types of graduation require-

ments significantly increase the probability of speaking a foreign language. High school graduation

requirements are the most effective, increasing the probability that the individuals speak a foreign

language by more than 18 percent.

Table 9 presents the estimates of the labor market returns to speaking a foreign language using

college and high school graduation requirements as instruments for the ability to speak a foreign

language. We estimate alternative specifications, including in some of them controls for other high

school requirements and college quality. Regardless of the specification, the conclusions are similar.

The use of graduation requirements as exogenous determinants of whether an individual speaks

a foreign language yields estimates of the returns to speaking a foreign language between 14 and

30 percent, much higher than the OLS estimates. Standard errors are large, as well, so that we

can not reject the hypothesis that the difference between the OLS and the IV estimators is zero at

conventional significance levels.25

24 The number of semesters in English, Social Studies, Math, and Science. See Table A.4 in the Appendix for acomprehensive list of state requirements.25 We experimented with alternative sets of instrumental variables in unreported regressions. One particular

concern is that college graduation requirements may reflect unobserved college quality and selectivity. It is thusreassuring that when we focus only on state high school graduation requirements our results do not change. We also

20

We estimate alternative IV models to test the robustness of our findings. As we did for the

OLS estimates, we restrict the sample to individuals who spoke English at home while growing up

(column 2), to American citizens who spoke English at home and whose parents were born in the

U.S. (column 3), and to full-time workers who spoke English at home (column 4). Again, the point

estimates remain stable across specifications, although they are large and very imprecise.

How can we interpret the differences between the returns to speaking a second language that

OLS regressions and propensity score methods suggest and our IV estimates? If we believe that

the “true” rate of return is homogeneous across the population, the IV estimates that rely on

exogenous sources of variation in the ability to speak a second language indicate that conventional

OLS estimates are downward biased. However, our preferred interpretation is the following: if the

rate of return to speaking a foreign language does vary across the population, and the population

that is affected by the graduation requirements (the compliers that we previously mentioned) have

high returns to learning a foreign language, the IV estimate, which depends on the marginal return

for this subset of the population, will be relatively high.26

6 Conclusions

The popular press has repeatedly presented anecdotal evidence suggesting that speaking a foreign

language is a valuable skill. However, little systematic research has addressed this issue. In this

paper we have explored this question using a variety of empirical strategies. While none of them

provides an ideal approach to the problem, all of them point in the same direction: speaking a foreign

language is rewarded in the labor market. The earnings of those who speak a foreign language are

higher than the earnings of those who don’t. This earnings differential does not seem to be driven

by unobservable differences between these two groups of individuals, as our IV estimates show. The

returns are not homogeneous across the population but are nevertheless positive and significant.

Learning a foreign language is an investment with positive monetary returns for the average college

graduate in the U.S.

This result could be valuable for policy-making purposes. In particular, frequent debates frame

the decisions to define high school curricula. Our results confirm that speaking a second language is

included the graduation requirements by state colleges, the changes in high school graduation requirements over time,etc. The estimates remain around 20 percent and very imprecise.26 A possible explanation for the high point estimates that we obtain in the IV regressions is that the second

language indicator is measured with nonclassical measurement error, which would bias the magnitude of IV coefficientsupwards (Kane et al., 1999). Another possibility is that we may also be introducing some contaminating variation ifthe instruments reflect unobserved individual and/or family background characteristics.

21

rewarded in the labor market with higher average earnings. It maybe therefore advisable to require

high school students to show proficiency in a foreign language before they graduate.

We believe that further research should assess the reasons for the positive returns to speaking

a foreign language. It would be interesting to examine if speaking a foreign language is used as

a signal for other abilities, for example. Is this skill valued per se, or only when it is actually

necessary on the job? For example, is it more valuable in those occupations and industries in which

foreign relations or international trade play an important role? The extension of the analysis to

other population groups in the U.S. or to individuals in other countries could show whether the

returns that we estimate in this paper extend to other populations. All these questions deserve

future exploration.

22

References

[1] Altonji, Joseph G. (1995) “The Effects of High School Curriculum on Education and LaborMarket Outcomes,” Journal of Human Resources, 30(3): 409-38.

[2] Angrist, J., and Guido Imbens (1994) “Identification and Estimation of Local Average Treat-ment Effects,” Econometrica, 62(2): 467-476.

[3] Angrist, J., and Alan B. Krueger (1999) “Empirical Strategies in Labor Economics,” in Hand-book of Labor Economics, Volume 3A. O.Ashenfelter, and David Card, eds. Amsterdam; NewYork and Oxford: Elsevier Science, North-Holland.

[4] Becker, G. (1964) Human Capital. New York: Columbia University Press.

[5] Brod, R., and M. Lapointe (1989) “The Modern Language Association Survey of ForeignLanguage Entrance and Degree Requirements, 1987-88,” ADFL Bulletin, 20(2).

[6] Carliner, G. (1995) “The Language Ability of US immigrants: Assimilation and Cohort Ef-fects,” NBER Working Paper #5222.

[7] Chiswick, B.R. (1997) "Hebrew Language Usage: Determinants and Effects on EarningsAmong Immigrants in Israel," Journal of Population Economics, 11(2): 253-271.

[8] Chiswick, B.R., and P.W. Miller (1997a) “Language Skill Definition: A Study of LegalizedAliens,” Mimeo.

[9] Chiswick, B.R., and P.W. Miller (1997b) “Language and Earnings Among Immigrants in theUnited States,” Mimeo.

[10] Chiswick, B.R., and P.W. Miller (1998) “English Language Fluency Among Immigrants in theUnited States,” Research in Labor Economics, 17: 151-200.

[11] Chiswick, B.R., and P.W. Miller (2000) “The Complementarity of Language and Other HumanCapital: Immigrant Earnings in Canada,” Center for the Study of the Economy and theState,Graduate School of Business, University of Chicago, Working Paper No. 148.

[12] Christofides, L.N. and Swidinsky, R. (1998) “Bilingualism and Earnings: A Study Based on1971, 1981 and 1991 Census Data,” in New Canadian Perspectives. Economic Approaches toLanguage and Bilingualism, A.Breton, ed.

[13] Cook, V. (1997) “The Consequences of Bilingualism for Cognitive Processing,” in Tutorialsin Bilingualism: Psycholinguistic Perspectives, A.M. De Groot, and J.F. Kroll, eds. Mahwah,New Jersey: Lawrence Erlbaum Associates.

[14] Cooper, T.C. (1987) “Foreign Language Study and SAT-Verbal Scores,” The Modern LanguageJournal, 71 (4).

[15] Dehejia, R.V., and Sadek Wahba (1999) “Causal Effects in Non-experimental Studies: Reeval-uating the Evaluation of Training Programs,” Journal of the American Statistical Association,94(448): 1053-1062.

[16] Grenier, G., and F. Vaillancourt (1983) “An Economic Perspective on Learning a SecondLanguage,” Journal of Multilingual and Multicultural Development, 4(6).

23

[17] Griliches, Z. (1977) “Estimating the Returns to Schooling: Some Econometric Problems,”Econometrica, 45(1): 1-22.

[18] Heckman, James (1980) “Sample Selection Bias as a Specification Error,” in Female LaborSupply, J. Smith, ed. Princeton NJ: Princeton University Press.

[19] Hirano, K., Imbens, G.W. and Geert Ridder (2000) “Efficient Estimation of Average TreatmentEffects Using the Estimated Propensity Score,” NBER Working Paper No. T0251.

[20] Hoffman, C. (1991) An Introduction to Bilingualism. London: Longman.

[21] Horvitz, D.G. and D.J.Thompson (1952) "A Generalization of Sampling Without Replacementfrom a Finite Universe." Journal of the American Stastistical Association, 47: 663-685

[22] Hoxby, C. M., and B. T. Long (1999) “Explaining Rising Income and Wage Inequality Amongthe College Educated,” NBER Working Paper No. 6873.

[23] Kane, Thomas J., Cecilia E. Rouse, and Douglas Staiger (1999) “Estimating Returns to School-ing When Schooling Is Misreported,” NBER Working Paper.

[24] Lalonde, R. (1986) “Evaluating the Econometric Evaluations of Training Programs,” AmericanEconomic Review 76(4): 604-620.

[25] Lopez, M.H. (1999) “Does Speaking a Second Language Affect Labor Market Outcomes?Evidence from National Adult Literacy Survey of 1992,” mimeo: School of Public Affairs,University of Maryland.

[26] Mincer, J. (1974) Schooling, Experience and Earnings. New York: Columbia University Press.

[27] Olsen, S., and L.K. Brown (1992) “The Relationship Between High School Study of ForeignLanguage and ACT English and Mathematics Performance,” ADFL Bulletin, 23(3):47-50.

[28] Pendakur, K., and R.Pendakur (1998) “Speak and Ye Shall Receive: Language Knowledge asHuman Capital” in New Canadian Perspectives. Official Languages in Canada: Changing theLanguage Landscape, A.Breton, ed. Ottawa : Canadian Heritage.

[29] Ridler, N.B. and S.Pons-Ridler (1984) "Language Economics: ACase Study of French," Journalof Multilingual and Multicultural Development, 5,1: 57-63.

[30] Rosenbaum, P. (1984) “The Consequences of Adjustment for a Concomitant Variable ThatHas Been Affected by the Treatment,” Journal of the Royal Statistical Society, 147(5).

[31] Rosenbaum, P., and D. Rubin (1983) “The Central Role of the Propensity Score in Observa-tional Studies for Causal Effects,” Biometrika, 70: 41-55.

[32] Rosenbaum, P., and D. Rubin (1984) “Reducing Bias in Observational Studies Using theSubclassification on the Propensity Score,” Journal of the American Statistical Association,79: 516-524.

[33] Rubin, D. (1974) “Estimating Causal Effects of Treatments in Randomized and Non-Randomized Studies,” Journal of Educational Psychology, 66: 688-701.

[34] Rubin, D. (1977) “Assignment to a Treatment Group on the Basis of a Covariate,” Journal ofEducational Statistics, 2:1-26.

24

[35] Rubin, D. (1978) “Bayesian Inference for Causal Effects: The Role of Randomization,” TheAnnals of Statistics, 6: 34-58.

[36] Shapiro, D.M., and M. Stelcner (1997) “Language and Earnings in Quebec: Trends overTwenty Years, 1970-1990,” Canadian Public Policy, 23(2): 115-140.

25

A Data Appendix

A.1 Quality of Schooling Variables

This is a set of dummies that indicate the quality of the college. The classification was providedby Bridget Terry Long and is defined in detail in Hoxby and Long (1999). Colleges were groupedaccording to criteria taken from the 1997 edition of Barron’s Profiles of American Colleges. Thesecriteria include raw test scores, percentage of the class scoring over a certain level, and high schoolrank. Schools that had not been ranked were categorized according to these criteria. These haveremained fairly constant over the last 15 years, and thus the use of the 1997 list is valid. Scoreswere reported as nonrecentered. The SAT scores were converted into percentiles.

Quality Groupings Criteria

Category G.P.A. S.A.T. A.C.T. Sample Schools

Most Competitive 3.75 - 4.50 1280 - 1600 31 - 36 Princeton, Harvard, UVA

Highly Competitive 3.40 - 3.74 1120 - 1279 26 - 30 UC Berkeley, University of Michigan

Very Competitive 3.00 - 3.39 1000 - 1119 21 - 25 University of Connecticut

Competitive 2.50 - 2.99 850 - 999 16 - 20 University of Massachusetts

Less Competitive 2.00 - 2.49 650 - 849 12 - 15 University of Mississippi

Non Competitive 1.00 - 1.99 649 - 400 4 - 11 Ohio State University - Mansfield Campus

Source: Barron’s Educational Series, Profiles of American Colleges, 21st Edition (1997).

Notes: The categories were created referring to Barron’s criteria for school groupings, which are based on the char-

acteristics of the freshman class entering 1995-96 (before SAT scores were recentered).

26

B A Simple Model

The following simple model illustrates the relation between ability and the returns to learning asecond language. It provides a justification for the strategy followed in the regressions in Table 7.Assume that the decision on learning a second language (SL hereafter), which involves a period oftraining, is undertaken before full participation in the labor market. Assume, furthermore, thatthere are two kinds of individuals. A fraction � of the population has low costs of learning an SL,�� (which corresponds to high ability). The rest of the population has costs �� � Ability is notobservable by the econometrician. Assume that if the second language is learned, it will turn outto be useful (producing extra earnings �� ) with probability � and turn out not to be useful in thelabor market with probability (1 − �). Assume that there are no consumption advantages fromspeaking a foreign language. Labor market earnings are:

�� =

½�� +�− ���� if the individual speaks an SL and the SL realization is positive�− ���� in all other cases

High ability individuals will decide to learn the second language if:

� · �� ≥ ��

For low ability individuals, the condition is:

� · �� ≥ ��

Consider the case when:�� ≥ � · �� ≥ ��

Then, all high ability individuals learn a second language and low ability individuals do not.Assume that individuals can maintain their foreign language human capital after the realization

of the labor market shock with a very small cost �� High ability individuals for which the foreignlanguage turned out to be useful will decide to maintain their stock of foreign language capital.High ability individuals for which the second language did not turn to be useful will decide notto maintain it. Ex-post, only individuals with high ability and a positive realization will speak asecond language. The pool of those who do not speak an SL will be formed by low ability individualsand high ability individuals with negative labor market realizations.

We know that the expected return of learning a second language, corresponding to a completerandomization of the second language treatment, is equal to � · �� . We may try to estimate thereturns to learning a second language using the difference between the wages of those who speakan SL and the rest (�� −��). Let T be an indicator that takes value one if the individual speaksa second language and zero otherwise. The expectation of this estimator is:

� (�� −��) = �(�|�� = 1)−�(�|�� = 0) == [�� +�− ���]−

·(1− �)�

1− ��· (�− ���) +

(1− �)

1− ��· (�− ���)

¸=

= �� +�(1− �)

(1− ��)· [�� −��] � �� � ���

This estimate is too high for two reasons. First, there is an ability bias: individuals who speaka second language have higher average unmeasured ability. Second, there is a problem of selectionby earnings: those who speak a second language ex-post tend to have greater returns than the rest.

27

These biases suggest the use of the alternative estimator: � · (�� −��� ��), where ��� �� is thewage for those individuals who learned the SL and did not invest in its maintenance ex-post. Thisis an unbiased estimator of the ex-ante return to speaking an SL:

� (�� −��� ��) = �(�|�� = 1)−�(�|�����) = [�� +�− ���]− [�− ���] = ��

Multiplying by the fraction of people who speak a second language out of the total who everspoke it yields the ex-ante expected returns to learning a second language. From column (1) inTable 4 we can derive:

(�� −��� ��) = �029− �002 = �027

The fraction of people who speak a second language out of the fraction who ever spoke it isequal to the number of people who speak divided by the number of people who speak plus thenumber of people who forgot:

� =��

�� +��� ��

=3298

3298 + 1103= 0�75

Thus, our estimate of the ex-ante expected returns of learning an SL is:

� · (�� −��� ��) = 0�020

Note how this is a conservative estimate of the expected returns to learning an SL, as we areassuming that the returns for those who forgot the second language is zero. Furthermore, note thatthe fact that the coefficient for those who forgot a second language is close to zero suggests thatthe ability bias is small.

28

Table 1:Comparison of Individuals Who Speak and Do Not Speak Foreign Languages, 1997 Survey

Panel A: Descriptive Statistics

Variable: Speak Not Speak DifferenceWorking? .955 .967 -.012∗∗∗

(�207) (�177) (�004)

Log Hourly Wage 2.562 2.544 .018∗

(�432) (�414) (�009)

Speaks F.L. 1 0 −− −

Age 29.611 30.356 -.744∗∗∗

(6�083) (6�755) (�153)

Experience (Months) 42.185 43.842 -1.657∗∗∗

(12�342) (11�011) (�270)

Married .244 .337 -.093∗∗∗

(�429) (�472) (�010)

Black .090 .014 .075∗∗∗

(�286) (�119) (�004)

Male .445 .44 -.003(�498) (�497) (�011)

Normalized GPA 3.062 3.063 -.001(�492) (�502) (�011)

Public College .653 .689 -.035∗∗∗

(�475) (�462) (�011)

MBA .023 .031 -.008∗∗

(�149) (�175) (�003)

Ph.D. .043 .021 .021∗∗∗

(�204) (�146) (�004)

Other Master’s .269 .249 .020∗∗

(�443) (�432) (�010)

Observations: 2756 5184 7940(34.71%) (65.28%)

29

Panel B: SAT Scores

Quartile: Speak Not Speak Difference1 .170 .203 -.033∗∗∗

(�376) (�403) (�009)

2 .202 .227 -.025∗∗∗

(�401) (�419) (�009)

3 .217 .198 .018∗∗

(�412) (�398) (�009)

4 .219 .165 .054∗∗∗

(�413) (�371) (�009)

Obs: 2756 5184 7940

Panel C: Regional Distribution, State of Residence

Region: % Speak ObservationsNew England .41 485

Middle Atlantic .38 999

East North Central .26 1231

West North Central .27 689

South Atlantic .35 1658

East South Central .26 541

West South Central .32 840

Mountain .34 544

Pacific .41 1028

Notes:

1. Subsample of individuals who have hourly wages above $2.8 (1 percentile) and below $42.3 (99 percentile), who answer the

question on whether they speak a foreign language and with complete data on age, experience, gender, marital status, race,

state of residence, college GPA, and type of college attended.

2. Sample means weighted using sample weights.

3. ∗∗∗ statistically significant at the 1% level, ∗∗ at the 5% level, ∗ at the 10% level.

4. New England: CT, ME, MA, NH, RI, VT; Middle Atlantic: NJ, NY, PA; East North Central: IN, IL, MI, OH, WI; West

North Central: IA, KS, MN, MO, NB, ND, SD; South Atlantic: DE, DC, FL, GA, MD, NC, SC, VA, WV; East South Central:

AL, KY, MS, TN; West South Central: AR, LA, OK, TX; Mountain: AZ, CO, ID, NM, MT, UT, NV, WY; Pacific: AK, CA,

HI, OR, WA.

30

Table 2: Descriptive Statistics, Individuals Who Speak a Foreign Language, 1997 Survey

Panel A: Do They Speak More Than One Language?

Sample: All Women Men DifferenceSpeaks 1 F.L. .840 .841 .838 -.0006

(�366) (�365) (�367) (�014)

Speaks 2 F.L. .135 .135 .134 -.001(�342) (�342) (�341) (�013)

Speaks 3 F.L. .018 .019 .018 -.001(�136) (�137) (�134) (�005)

Speaks �3F.L. .005 .003 .007 .004(�076) (�056) (�088) (�010)

Observations: 2756 1530 1226 2756

Panel B: What Languages Do They Speak?

Sample: All Women Men DifferenceSpanish .58 .586 .589 .002

(�49) (�492) (�492) (�019)

French .23 .272 .181 -.090∗∗∗

(�42) (�445) (�385) (�016)

German .11 .081 .149 .067∗∗∗

(�31) (�273) (�356) (�012)

Italian .03 .043 .023 -.020∗∗∗

(�18) (�204) (�150) (�007)

Russian .01 .012 .021 .008∗

(�12) (�112) (�145) (�005)

E/SE Asian .04 .026 .062 .035∗∗∗

(�20) (�161) (�242) (�007)

Chinese .02 .024 .027 .002(�15) (�156) (�163) (�006)

Obs.: 2756 1530 1226 2756

Panel C: Regional Distribution

Middle E.N W. N. South E. S. W. S.Region: N. E. Atlantic Central Central Atlantic Central Central Mountain Pacific

Spanish .52 .53 .54 .45 .61 .67 .72 .65 .56

French .34 .25 .22 .17 .28 .25 .15 .14 .20

German .09 .10 .15 .16 .09 .08 .10 .08 .10

Italian .09 .08 .01 .008 .03 .02 .006 .008 .01

Russian .02 .01 .01 .008 .02 .006 .005 .04 .008

E/SE Asian .008 .05 .04 .16 .01 .009 .01 .01 .06

Chinese .03 .04 .01 .01 .01 .005 .02 .008 .04

Notes:

1. Subsample of individuals who have hourly wages above $2.8 (1 percentile) and below $42.3 (99 percentile), who answer YES

to the question on whether they speak a foreign language and with complete data on age, experience, gender, marital status,

race, state of residence, college GPA, and type of college attended.

2. See Notes 2, 3 and 4 to Table 1.

31

Table 3: OLS EstimatesDependent Variable: Log Hourly Wage in 1997

(1) (2) (3) (4)Spoke English at Home, Spoke English

Spoke English American Citizen, at Home,Sample: All at Home Parents Born in US Works�35h.

Speaks F.L. .028∗∗∗ .022∗∗ .022∗∗ .019∗

(.01) (�010) (�010) (�010)

Age .030∗∗ .035∗∗∗ .032∗∗∗ .031∗∗∗

(.006) (�008) (�006) (�006)

Age2 -.0003∗∗∗ -.0003∗∗∗ -.0003∗∗∗ -.0003∗∗∗

(�00007) (�0001) (�00008) (�00008)

Experience .002 .004 .0016 -.004∗∗

(�002) (�002) (�002) (�002)

Experience2 -.00001 -.00004 -.00001 .00005∗

(�00003) (�00003) (�00003) (�00003)

Male .071∗∗∗ .077∗∗∗ .074∗∗∗ .062∗∗∗

(�010) (�010) (�010) (�010)

Married .040∗∗∗ .041∗∗∗ .048∗∗∗ .034∗∗∗

(�010) (�011) (�011) (�011)

Black .061∗∗∗ .005 .020 -.001(�023) (�032) (�055) (�031)

Log State Income .515∗∗∗ .449∗∗∗ .494∗∗∗ .442∗∗∗

(�034) (�054) (�037) (�037)

Public College .004 .006 -.002 -.007(�009) (�011) (�010) (�010)

Normalized .015∗ .008 .011 -.001College GPA (�008) (�009) (�009) (�009)

Parents’ Education Yes Yes Yes Yes

College Quality Yes Yes Yes Yes

Graduate Degree Yes Yes Yes Yes

SAT-ACT Quartile Yes Yes Yes Yes

Major Yes Yes Yes Yes

Adj. R2 .186 .182 .182 .198Observations: 7940 7073 6629 5906

Notes:

1. Subsample of individuals who have hourly wages above $2.8 (1 percentile) and below $42.3 (99 percentile). Each regression

is performed for the maximum number of observations for which all the covariates were non-missing.

2. Observations are weighted using sample weights.

3. Standard errors are reported in parenthesis.

4. Log State Income is the log of the per capita income in the state of residence in 1997 (BEA estimation).

5. Normalized GPA on a 0-4 scale for all respondents.

6. Parents’ Education is a set of dummies that capture the education level of the individual’s mother and father.

7. Majors is a detailed set of indicators for the student’s major in college (100 categories).

8. Quality of College Dummies were provided by Bridget Terry Long. See the Appendix for a description, and Hoxby and Long,

(1999) for details.

9. Graduate Degree dummies are three variables that indicate if the individual has a Ph.D., an MBA, or a Master’s.

10. ∗∗∗ statistically significant at the 1% level, ∗∗ at the 5% level, ∗ at the 10% level.

32

Table 4: A Test on Ability BiasDependent Variable: Log Hourly Wage in 1997

(1) (2) (3)Spoke English at Home,

Spoke English at Home, American Citizen,Sample: All American Citizen Parents Born in US.

Speaks FL in 1997 .029∗∗∗ .024∗∗ .024∗∗

(�010) (�011) (�011)

Spoke FL in 1993 only .002 .007 .005(�014) (�015) (�015)

Age .029∗∗∗ .030∗∗∗ .031∗∗∗

(�006) (�008) (�006)

Age2 —.0003∗∗∗ -.0003∗∗∗ -.00031∗∗∗

(�0001) (�00008) (�00001)

Experience .003 .003 .002(�002) (�002) (�002)

Experience2 -.00002 -.00003 -.00002(�00003) (�00003) (�00003)

Male .074∗∗∗ .081∗∗∗ .075∗∗∗

(�010) (�010) (�010)

Married .039∗∗∗ .050∗∗∗ .049∗∗∗

(�011) (�011) (�011)

Black .056∗∗ .014 .019(�023) (�032) (�036)

Log State Income .525∗∗∗ .521∗∗∗ .497∗∗∗

(�053) (�036) (�086)

Public College .005 .004 -.0002(�010) (�010) (�010)

Normalized .017∗ .011 .010College GPA (�009) (�009) (�009)

College Quality Yes Yes Yes

SAT-ACT Quartile Yes Yes Yes

Major Yes Yes Yes

Parents’ Education Yes Yes Yes

Graduate Degree Yes Yes Yes

Adjusted R2 .188 .182 .184Observations: 7444 6632 6117

Notes:

1. Sample of individuals who have hourly wages above the 1st percentile and below the 99th percentile, who answer the question

of whether they speak a foreign language in both surveys, and for whom the variables included in the regressions were complete.

2. Observations are weighted using sample weights.

3. Standard errors in parenthesis.

4. See notes to Table 3 for the variable definition.

5. ∗∗∗ statistically significant at the 1% level, ∗∗ at the 5% level, ∗ at the 10% level.

33

Table 5: Estimated Impact of Speaking a Second LanguagePropensity Score Methods

Panel A. Stratifying on the Score

Not Adjusted AdjustedLog.Wage Diff. #Treated #Control Log.Wage Diff. #Treated #Control

Decile 1 .005 (�037) 173 603 .009 (.035) 166 574

Decile 2 -.033 (�034) 207 585 .027 (.033) 195 556

Decile 3 -.029 (�034) 194 585 -.014 (.033) 181 559

Decile 4 .063∗ (�035) 194 575 .036 (.033) 188 550

Decile 5 -.006 (�032) 239 544 .036 (.031) 229 522

Decile 6 -.022 (�031) 271 505 .024 (.030) 259 480

Decile 7 -.025 (�032) 253 515 .000 (.031) 240 500

Decile 8 -.087∗∗∗ (�031) 299 481 -.022 (.029) 287 459

Decile 9 .024 (�031) 314 434 .051∗ (.029) 304 414

Decile 10 .118∗∗∗ (�034) 518 200 .036 (.036) 483 190

Observations: 2662 5027 2532 4804b���� ������ .012 (�011) b�����

�� .020∗∗ (�010)

B. Matching on the Score

Unadjusted AdjustedSpeaks FL .028∗∗∗ (.011) .040∗∗∗ (�011)

R2 .001 .186Observations: 5559

Notes:

1. Sample of individuals who have hourly wages above the 1st percentile and below the 99th percentile, who answer the question

of whether they speak a foreign language in both surveys, and for whom the variables included in the regressions were complete.

In Panel B, the sample is further restricted to individuals who speak a foreign language and to those who don’t with the closest

propensity score.

2. Propensity scores are estimated using the logistic model presented in Table A.3.

3. Adjusted results include the set of controls presented in Table 3.

4. Standard errors in parenthesis.

5. ∗∗∗ statistically significant at the 1% level, ∗∗ at the 5% level, ∗ at the 10% level.

34

Table 6: Exploiting Information from First SurveyDependent Variable: Log Hourly Wage 1997-Log Hourly Wage 1994

(1) (2) (3) (4) (5)Spoke English at Home, Spoke English at Home,American Citizen, American Citizen,

Sample: All All Parents Born in US All Parents Born in US

∆Speaks FL .017∗ .017∗ .022∗∗ .012 .017∗

(�009) (�009) (�009) (�009) (�009)

Speaks F.L. - -.003 -0.011 .002 -.009both in 1997 and 1993 (�009) (�011) (�010) (�011)

∆Age2 .0012∗∗∗ .0012∗∗∗ .001∗∗∗ .002∗∗ .003∗∗∗

(�0001) (�0001) (�0001) (�0008) (�0009)

Experience -.002 -.002 -.001 .0001 -.0003(�002) (�001) (�002) (�0017) (�0020)

Experience2 .00005∗∗ .00005∗∗ .00005∗ .00002 .00003(�00002) (�00002) (�00003) (�00002) (�00003)

∆Log State Income -.017 -.017 -.011 .015 .006(�051) (�051) (�055) (�054) (�059)

∆Married .007 .007 .021∗∗ .012 .026∗∗

(�009) (�009) (�010) (�010) (�011)(Age in 1994)2 - - - -.00008 -.0002

(�00007) (�00008)

Male - - - .053∗∗∗ .057∗∗∗

(�008) (�009)

Married in 1994 - - - .017 .013(�010) (�011)

Black - - - -.037∗ .001(�021) (�054)

Log State Income 1994 - - - .128∗∗∗ .126∗∗

(�034) (�037)

Public College - - - -.039∗∗∗ -.036∗∗∗

(�009) (�010)

Normalized - - - -.001 -.001College GPA (�003) (�003)

College Quality No No No Yes Yes

SAT-ACT Quartile No No No Yes Yes

Major No No No Yes Yes

Parents’ Education No No No Yes Yes

Graduate Degree Yes Yes Yes Yes Yes

Adjusted R2 .032 .031 .028 .092 .092Observations: 7686 7686 6329 7248 5977

Notes: See Notes to Table 3.

35

Table 7: Exploiting Information from First SurveyAllows Asymmetry between Returns to Learning and ForgettingDependent Variable: Log Hourly Wage 1997−Log Hourly Wage 1994

(1) (2) (3) (4) (5)Spoke English at Home, Spoke English at Home,American Citizen, American Citizen,

Sample: All All Parents Born in US All Parents Born in US

Speaks FL 1997 only .037∗∗ .038∗∗ .028∗ .035∗∗ .026(�015) (�015) (�016) (�016) (�096)

Spoke FL 1993 only -.003 -.003 -.017 .004 -.007(�012) (�012) (�013) (�013) (�013)

Speaks FL - .002 -0.009 .008 -.007both in 1997 and 1993 (�009) (�011) (�010) (�011)

∆Age2 .001∗∗∗ .001∗∗∗ .001∗∗∗ .002∗∗ .003∗∗∗

(�0001) (�0001) (�0001) (�0008) (�0009)Experience -.002 -.002 -.002 .00007 -.0002

(�002) (�002) (�002) (�002) (�002)

Experience2 .00005∗∗ .00005∗∗ .00005∗ .00002 -.00003(�00002) (�00002) (�00003) (�00002) (�00003)

∆Log State Income -.015 -.011 -.011 .016 -.007(�051) (�051) (�055) (�054) (�06)

∆Married .007 .007 .021∗∗ .012 .026∗∗

(�009) (�009) (�010) (�010) (�011)Age2 - - - -.00008 -.0002∗∗

(�00007) (�00007)

Male - - - .053∗∗∗ .056∗∗∗

(�008) (�009)

Married - - - .018∗ .014(�010) (�011)

Black - - - -.037∗ .002(�021) (�054)

Log State Income - - - .128∗∗∗ .126∗∗∗

(�034) (�037)

Public College - - - -.038∗∗∗ -.04∗∗∗

(�009) (�010)

Normalized - - - -.001 -.001College GPA (�003) (�003)

College Quality No No No Yes Yes

SAT-ACT Quartile No No No Yes Yes

Major Dummies No No No Yes Yes

Parents’ Education No No No Yes Yes

Graduate Degree Yes Yes Yes Yes Yes

Adjusted R2 .032 .032 0.027 0.073 0.069Observations: 7686 7686 6329 7248 5977

Notes: See Notes to Table 3.

36

Table 8: Instrumental Variable EstimatesFirst Stage

Dependent Variable: Speaks Foreign Language in 1997Linear Probability Model

(1) (2)HS Requirement .188∗∗∗ .212∗∗∗

(�049) (�050)

HS Elective .072∗∗∗ .074∗∗∗

(�021) (�023)

College Requirement .024∗ .031∗∗

(�013) (�014)

HS Requirement·College Req. -.125∗∗ -.119∗∗

(�054) (�053)

HS Elective·College Req. .004 -.043(�028) (�028)

Other Controls? No Yes

R2 .006 .127Observations: 7705 7179

Notes:

1. Rhode Island, Texas, and the District of Columbia required high school students to take foreign language courses; California,

New Hampshire, Oklahoma, Oregon, Virginia, and West Virginia included foreign language courses among the elective courses

needed to fulfill graduation requirements in 1989. A complete list of college graduation requirements is available upon request

from the authors.

2. Column 2 includes controls for age, age squared, experience and experience squared, marital status (married or not), race,

the log of the average income in the state of residence, an indicator of attending a public college, normalized GPA, the quality

of the college the individual attended, major, parents education level, whether the individual holds a graduate degree, and other

high school requirements.

3. See Notes to Table 3 for a sample and variable description.

37

Table 9: Instrumental Variable EstimatesDependent Variable: Log Hourly Wage in 1997

(1) (2) (3) (4)Spoke English at Home, Spoke English

Spoke English American Citizen, at Home,Sample: All at Home Parents Born in US Works�35h.

Speaks F.L. .271∗ .198 .214 .169(�147) (�174) (�184) (�182)

Age .031∗∗∗ .035∗∗∗ .035∗∗∗ .035∗∗∗

(�008) (�008) (�008) (�009)

Age2 -.0003∗∗∗ -.0003∗∗∗ -.0003∗∗∗ -.0003∗∗∗

(�0001) (�001) (�002) (�0001)

Experience .003 .003 .004∗ -.002(�002) (�002) (�002) (�002)

Experience2 -.00003 -.00003 -.00004 .0003(�00003) (�00003) (�00003) (�00003)

Male .072∗∗∗ .075∗∗∗ .073∗∗∗ .057∗∗∗

(�010) (�010) (�011) (�010)

Married .053∗∗∗ .050∗∗∗ .058∗∗∗ .039∗∗∗

(�013) (�013) (�015) (�013)

Black -.037 -.042 .004 -.044(�059) (�054) (�061) (�060)

Log State .463∗∗∗ .469∗∗∗ .453∗∗∗ .408∗∗∗

Income (�041) (�039) (�040) (�040)

Normalized .008 .005 .005 .003College GPA (�009) (�003) (�004) (�003)

Public College .008 .003 -.004 -.006(�010) (�011) (�011) (�011)

Parents’ Education Yes Yes Yes Yes

Major Yes Yes Yes Yes

SAT-ACTQuartile Yes Yes Yes Yes

College Quality Yes Yes Yes Yes

Graduate Degree Yes Yes Yes Yes

Other HS Req. Yes Yes Yes Yes

Observations: 7179 6775 6231 5674

Notes:

1. Instruments: State level foreign language requirement in high school in 1989, foreign language requirement among electives

in high school in 1989 at the state level, indicators that college the individual attended had a foreign language requirement,

and interactions of the previous dummy variables. Rhode Island, Texas and the District of Columbia required high school

students to take foreign language courses; California, New Hampshire, Oklahoma, Oregon, Virginia and West Virginia included

foreign language courses among the elective courses needed to fulfill graduation requirements in 1989. A complete list of college

graduation requirements is available upon request from the authors.

2. Other High School Requirements is a set of variables that capture the requirements in the state where the individual studied

in High School. They include English, Social Studies, Math, and Science. See Appendix Table A.5 for a complete description

of these requirements.

3. See Notes to Table 3 for a sample and variable description.

38

Table A.1: Summary StatisticsBachelor and Beyond Sample, 1997 Survey

Panel A: Demographics

Variable: Mean Standard Dev. Min. MaxLog Hourly Wage 2.55 .420 1.03 3.73

Speaks FL .340 .474 0 1

Age 30.102 6.543 18 73

Experience 43.277 11.508 0 57

Married .305 .460 0 1

Black .040 .196 0 1

Male .448 .497 0 1

Normalized GPA 3.063 .498 1.17 4

Public College .677 .467 0 1

MBA .028 .167 0 1

Ph.D. .029 .168 0 1

Other Masters .256 .436 0 1

Observations: 7940

Panel B: Regional Distribution of Observations(State of Residence)

Region:New England .10

Middle Atlantic .12

East North Central .15

West North Central .08

South Atlantic .20

East South Central .06

West South Central .11

Mountain .06

Pacific .12

Notes:

1. Subsample of Individuals who have hourly wages above $2.8 (1 percentile) and below $42.3 (99 percentile), who answer the

question on whether they speak a foreign language and with complete data on age, experience, gender, marital status, race,

state of residence, college GPA, and type of college attended.

2. Mean values weighted using sample weights.

3. New England: CT, ME, MA, NH, RI, VT; Middle Atlantic: NJ, NY, PA; East North Central: IN, IL, MI, OH, WI; West

North Central: IA, KS, MN, MO, NB, ND, SD; South Atlantic: DE, DC, FL, GA, MD, NC, SC, VA, WV; East South Central:

AL, KY, MS, TN; West South Central: AR, LA, OK, TX; Mountain: AZ, CO, ID, NM, MT, UT, NV, WY; Pacific: AK, CA,

HI, OR, WA.

39

Table A.2: Percentage of Second Language Speakers and Average Earnings by Major, 1997Survey% Speak a FL Average Log(Wage)

Health .214 (.021) 2.818 (.019)

Engineering .216 (.022) 2.844 (.020)

Computer Sciences .219 (.036) 2.774 (.032)

Vocational/Technical .226 (.036) 2.524 (.032)

Business .227 (�011) 2.599 (�010)

Education .271 (�014) 2.384 (�012)

Life Sciences .277 (�021) 2.404 (�019)

Other Technical/Professional .318 (�018) 2.530 (�016)

Social Sciences .330 (�014) 2.489 (�012)

Mathematics .346 (�044) 2.519 (�039)

Physical Sciences .382 (�047) 2.496 (�041)

Humanities .450 (�017) 2.428 (�015)

40

Table A.3: Estimated Effect of Speaking a Foreign LanguagePropensity Score Estimation

Logit Model

Dependent Variable: Speaks F.L. in 1997

CoefficientAge -.103∗∗ (.041)Age2 -.001∗∗ (.0005)Male .006 (.052)

Black .353∗∗ (.162)Log State Income .650∗∗∗ (.165)Spoke English at Home -1.848∗∗∗ (.144)Mother Born in US -1.059∗∗∗ (.102)Mother’s Education Yes

Father’s Education Yes

Observations: 7684

Notes:

1. Observations weighted using sample weights.

2. ∗∗∗ statistically significant at the 1% level, ∗∗ at the 5% level, ∗ at the 10% level.

41

Table A.4: Descriptive Statistics: Pre-Determined CovariatesMatched Sample

Sample: Speak Not Speak DifferenceAge 29.648 29.39 .249

(6�142) (5�61) (�166)

Black .088 .071 .017(�283) (�257) (�007)

Male .444 .461 -.016(�496) (�498) (�014)

Log State Income 9.80 9.80 .005(�157) (�148) (�004)

Spoke English at Home .770 .908 -.137∗∗∗

(�420) (�288) (�009)

Mother Born in US .758 .870 -.111∗∗∗

(�427) (�335) (�010)

Observations: 3097 2451 5548

Notes:

1. Sample of individuals who have hourly wages above $2.8 (1 percentile) and below $42.3 (99 percentile), and who answer

the question of whether they speak a foreign language. Each individual who speaks a foreign language was matched with

an observation from the subsample of individuals who don’t speak a foreign language with the closest propensity score. The

matching was with replacement (i.e., each control observation was allowed to be the match for more than one treated observation).

2. Mean values weighted using sample weights, taking into account that the matching was with replacement.

42

Table A.5: High School Graduation Requirements in 1989

F.L F.L. Social PhysicalState: Req. Elec. All English Studies Math Science Ed. ElectivesAlabama 0 0 20 4 3 2 1 3.5 6.5Alaska 0 0 21 4 3 2 2 1 9Arizona 0 0 20 4 3 3 3 1 6.5Arkansas 0 0 20 4 3 3 3 1 6.5California 0 1 13 3 3 2.5 2.5 2 -Colorado1 0 0 - - - - - - -Connecticut 0 0 20 4 3 3 2 1 6Delaware 0 0 19 4 3 2 2 1.5 6.5D.C.2 1 0 20.5 4 2 2 2 1.5 8Florida 0 0 24 4 3 3 3 1 9Georgia 0 0 21 4 3 2 2 1 8Hawaii 0 0 20 4 4 2 2 1.5 6Idaho 0 0 20 4 2 2 2 1.5 6Illinois 0 0 16 3 2 2 1 4.5 2.25Indiana 0 0 19.5 4 2 2 2 1.5 8Iowa3 0 0 - - 1.5 - - 1 -Kansas 0 0 20 4 3 2 2 1 8Kentucky 0 0 20 4 2 3 2 1 7Louisiana 0 0 23 4 3 3 3 2 7.5Maine 0 0 16 4 2 2 2 1.5 3.5Maryland 0 0 20 4 3 3 2 1 5Massachusetts3 0 0 - - 1 - - 4 -Michigan3 0 0 - - 0.5 - - - -Minnesota 0 0 20 4 3 1 1 1.5 9.5Mississippi 0 0 16 3 2.5 1 1 0 8.5Missouri4 0 0 24 4 3 3 3 1 8Montana 0 0 20 4 1.5-2 2 1 1 10.5-10Nebraska5 0 0 - - - - - - -

43

Table A.5, Continued: High School Graduation Requirements in 1989

F.L F.L. Social PhysicalState: Req. Elec. All English Studies Math Science Ed. ElectivesNevada 0 0 20 3 2 2 1 2.5 9.5

New Hampshire 0 1 19.75 4 2.5 2 2 1.25 4New Jersey 0 0 18.5 4 2 2 1 4 4New Mexico 0 0 21 4 2 2 2 1 9New York 0 0 18.5 4 4 2 3 0.5 varies

North Carolina 0 0 20 4 2 2 2 1 9North Dakota 0 0 17 4 3 2 2 1 5

Ohio 0 0 18 3 2 2 1 1 9Oklahoma4 0 2 15 4 2 3 2 0 4Oregon 0 1 22 3 3.5 2 2 2 8

Pennsylvania 0 0 21 4 3 3 3 1 2Rhode Island6 2 0 18 4 2 3 2 0 4South Carolina 0 0 20 4 3 3 2 1 7South Dakota 0 0 20 4 3 2 2 0 8Tennessee 0 0 20 4 1.5 2 2 1.5 9Texas 2 0 21 4 2.5 3 2 2 7Utah 0 0 24 3 3 2 2 2 9

Vermont 0 0 15.5 4 3 3 3 1.5 0Virginia 0 3 20 4 3 2 2 2 6

Washington 0 0 18 3 2.5 2 2 2 5.5West Virginia 0 1 21 4 3 2 1 2 7Wisconsin 0 0 13.5 4 3 2 2 2 0.5Wyoming 0 0 18 1 - - - - -

Source: Digest of Education Statistics, 1985-1986. Office of Educational Research and Improvement, U.S. Department of

Education Center for Statistics.

Notes:1Local boards determine requirements. The state has constitutional prohibition against state requirements.2For comprehensive diploma.3Local Boards determine additional requirements.4For college preparatory studies certificate.5200 credit hours required, at least 80 percent in core curriculum courses. The state was conducting hearings to define core

courses at the time of the survey.6College bound degree.

44

Table A.6: Descriptive Statistics

Panel A: Comparison of Individuals Whose Parents Live in 1993 in States with and Without ForeignLanguage High School Graduation Requirements

Variable: Req. No Req. DifferenceLog State Income 9.808 9.779 .029∗∗∗

(�117) (�170) (�003)

Log Hourly Wage 2.571 2.544 .027∗∗

(�417) (�421) (�011)

Speaks FL .398 .323 .074∗∗∗

(�489) (�467) (�012)

Age 29.7 30.224 -.524∗∗∗

(5�626) (6�793) (�172)

Experience 43.273 43.278 -.004(11�322) (11�565) (�304)

Married .263 .318 -.054∗∗∗

(�440) (�465) (�012)

Black .065 .032 .032∗∗∗

(�247) (�177) (�005)

Male .446 .448 -.001(�497) (�497) (�013)

Normalized 3.048 3.067 -.019College GPA (�506) (�496) (�013)

Public College .716 .665 .051∗∗∗

(�450) (�472) (�012)

MBA .026 .029 -.002(�161) (�169) (�004)

Ph.D. .034 .027 .006(�182) (�164) (�004)

Other Master’s .257 .255 .001(�437) (�436) (�011)

Observations: 1855 6085 7940

Notes: See Notes to Table 1.

45

Panel B.1: Comparison of Individuals by Existence of Requirement in College Attended

Variable: Req. No Req. DifferenceLog State Income 9.929 9.956 -.027∗∗∗

(�135) (�129) (�003)

Speaks F.L. .344 .332 .012(�475) (�471) (�011)

Age 29.566 30.185 -1.61∗∗∗

(5�889) (7�584) (�154)

Experience 43.135 43.563 -.427(11�702) (11�103) (�273)

Married .291 .334 -.042∗∗∗

(�454) (�471) (�010)

Black .035 .050 -.015∗∗∗

(�183) (�219) (�004)

Male .452 .439 .013(�454) (�496) (�011)

Normalized 3.055 3.078 -.023∗∗

College GPA (�493) (�509) (�011)

Public College .752 .524 .227∗∗∗

(�431) (�499) (�010)

MBA .032 .020 .011∗∗∗

(�178) (�143) (�003)

Ph.D. .032 .023 .004∗∗

(�178) (�151) (�004)

Other Masters .252 .263 -.010(�434) (�440) (�010)

Observations: 5211 2729 8624

Panel B.2: College Quality and Foreign Language Requirements

Category: Req. #Colleges # Individuals in SampleMost Competitive .684 19 315

(�477)Highly Competitive .454 33 549

(�505)Very Competitive .745 102 1642

(�437)Competitive .613 251 3198

(�490)Less Competitive .5 96 940

(�502)Non Competitive .386 75 562

(�490)

Public .614 324 4976(�487)

Private .464 278 2558(�499)

46